docs(e2e): Chapter 2 exploration runbook — kick-off, triage, promotion (#401)

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.

[AndresL230]

Chapter 2 of the E2E program (epic #403), PR 3 of 3. Closes#401.

The operator's manual for make explore / /explore (PR #444) and the oracle CLI (PR #443): what Chapter 2 is (vs Chapter 1's scripted journeys — local-only, never gates), prerequisites, both kick-off modes with the full knob table, the .explore/ output inventory, standalone oracle usage, the triage protocol for findings.md (reproduce → issue with evidence; noise → tune the oracle; known-bug → evidence only if new), the promotion pipeline (issue → scripted Chapter 1 journey, testids per docs/frontend-testids.md; epic bar: ≥3 findings promoted into Chapter 1 regressions in the first month), cadence/cost, the machine-singleton lock protocol, and troubleshooting.

Written against the MERGED harness (post-review reality): unconditional dummy GEMINI_API_KEY + SEED_RICH=1, Edit(.explore/**)-scoped explorer writes, EXPLORE_USER five-user wiring, holder-written lock.ok bookkeeping, session recordings under .explore/traces/. Every command, knob, default, path, and exit code in the doc was verified against the code (knob table cross-checked with /usr/bin/grep -o 'EXPLORE_[A-Z_]*', check names against e2e_oracles.__main__.CHECKS, make -n targets).

Plus one pointer line in docs/local-supabase.md's "One-command E2E stack" section.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Documentation
  • Added a comprehensive guide for AI-driven exploratory end-to-end testing.
  • Documented headless and interactive test runs, prerequisites, generated artifacts, oracle usage, troubleshooting, and issue triage.
  • Added guidance for promoting exploratory findings into regression coverage.
  • Linked the new exploratory testing guide from the local Supabase documentation.

[cloudflare-workers-and-pages]

Deploying with Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	frontend-staging	df2672b	Commit Preview URL Branch Preview URL	Jul 28 2026, 10:57 AM

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 202a3c9c-88e8-431a-a2d8-e3646f439c01

📥 Commits

Reviewing files that changed from the base of the PR and between 9f64d66 and df2672b.

📒 Files selected for processing (2)

• docs/e2e-exploration.md
• docs/local-supabase.md

---

📝 Walkthrough

Walkthrough

Adds a complete Chapter 2 exploratory-testing runbook covering execution modes, prerequisites, artifacts, oracle validation, finding triage, regression promotion, locking, cadence, and troubleshooting. Links the runbook from the local Supabase documentation.

Changes

Exploratory testing runbook

Layer / File(s)	Summary
Exploration execution and outputs docs/e2e-exploration.md	Documents prerequisites, headless and interactive workflows, generated artifacts, and stack startup and teardown behavior.
Oracle validation and result interpretation docs/e2e-exploration.md	Describes standalone oracle commands, selective checks, seeded users, stack assumptions, and exit codes.
Finding triage and regression promotion docs/e2e-exploration.md	Defines reproduction-based triage and promotion of confirmed findings into scripted Chapter 1 coverage.
Operational procedures and documentation entry point docs/e2e-exploration.md, docs/local-supabase.md	Adds cadence, cost, locking, troubleshooting, and a local Supabase link to the runbook.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related issues

• #403 — The runbook documents the Chapter 2 AI Agent Explorer workflow, oracle pipeline, and finding-promotion process described by this epic.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch e2e-chapter2-runbook

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[AndresL230]

Code review

Found 1 issue:

1. The triage section's flagship example presents a "wrong-data" upload bug (a newly uploaded document coming back with another document's summary/category/concepts) as a genuine promotable finding — but Chapter 2 always runs under SAPLING_MODEL_MODE=function with agents.function_handlers_e2e, whose classifier/summary/concepts handlers return the identical canned output ("Gradient Descent" abstract, lecture_notes category) for every upload by documented design ("Fixed output. Replies are constants…"). The observed symptom is the deterministic seam working as designed — precisely the "noise the harness produced" category the doc's own triage protocol says to drop — not a data-leak bug. Enshrining it as the worked example will teach operators to misfile seam artifacts. (The tutor-resume 500 from the same run — concept_describe unregistered — remains a genuine example and a better candidate for this slot.)

Sapling/docs/e2e-exploration.md

Lines 175 to 190 in 3c65b97

	Real evidence for what a genuine finding looks like: this harness's own
	bounded acceptance testing turned up a novel "wrong-data" bug (a newly
	uploaded, unrelated document came back with another document's summary,
	category, and extracted concepts verbatim — polluting the graph with wrong
	concepts) and, in a separate run of the same acceptance round, a real
	tutor-resume `500` on resuming a session (`POST
	/api/graph/<user>/concept-description` failing with `LookupError: ... no
	handler is registered for task 'concept_describe'` — `concept_describe` is
	genuinely unregistered in `backend/agents/function_handlers_e2e.py`),
	independently corroborated by the `logscan` oracle picking up the same `500`
	and traceback. Neither of those is on the known-bugs list — they're exactly
	the kind of thing this chapter is for. By contrast, #355 (the graph's
	duplicated CS subject-root hub) shows up in the graph oracle on essentially
	every run against the rich seed data — that's the known-not-new case
	category 3 above exists for.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[AndresL230]

Review finding fixed in 56511af: the promotable worked example is now the genuine tutor-resume 500 (concept_describe unregistered — facts re-verified against _providers.py/graph.py/main.py by the re-review), and the wrong-data episode is recast as a seam-artifact triage lesson with the recognition tell. Timing contradiction and dangling cross-refs fixed in the same commit; a code-span rewrap nit fixed in the follow-up. Scoped re-review: all findings addressed.