test: assert the real OPENCODE_CONFIG_CONTENT injection + merge (close boot_smoke false-green) by Rchari1 · Pull Request #55 · harmoniqs/amicode

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.

Rchari1 · 2026-07-01T16:22:36Z

Closes #25.

boot_smoke.mjs is the only server-booting test and it boots withoutOPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely — a false-green surfaced in #12 review.

Where the fix lives

The real fix must use the realbuildOpencodeConfigContent (a transcribed copy would just drift — the thing #25 warns about). A .mjs boot script can't import the TS builder, so the assertion belongs in opencode_config.test.ts, which already runs the real binary via opencode debug config (resolves the merged config ≈ GET /config) with the real builder.

Change

test/opencode_config.test.ts — hardened the real-binary merge test to a full injection+merge assertion. It now writes a user global config with a distinctive model + permission key and asserts:
- the injected instructions (the AGENTS.md path) is present — the exact regression the false-green missed;
- the user's global model survives the deep-merge (opencode's provider selection, Q129);
- the user's permission key survives and our injected external_directory is added (the fix: chat→solve QA — permission-hang blocker + multi-qubit/regime guidance #22 deep-merge check, folded in).
test/boot_smoke.mjs — documented its scope: binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts.

Validation

Full suite green (80 pass); test:smoke PASS.
Genuine catcher: temporarily dropping instructions from the builder reds this test at the instructions assertion (and the unit test) — proving it now catches the regression the false-green allowed.

🤖 Generated with Claude Code

…(close boot_smoke false-green) boot_smoke.mjs is the only server-booting test and boots WITHOUT OPENCODE_CONFIG_CONTENT, so it stayed green even if the instructions/permission injection were removed entirely (false-green surfaced in #12 review). The real fix belongs where the REAL buildOpencodeConfigContent can be imported (a .mjs can't import the TS builder without re-deriving it → drift). So harden the existing real-binary test in opencode_config.test.ts: it already runs `opencode debug config` (resolves the merged config, ≈ GET /config) with the real builder; now it also - writes a user global config with a distinctive `model` + `permission` key, - asserts the injected `instructions` (AGENTS.md path) is present — the exact regression the false-green missed, - asserts the global `model` survives the deep-merge (Q129 provider selection), - keeps the #22 permission deep-merge assertions (user key survives + ours added). Verified it's a real catcher: dropping `instructions` from the builder reds this test (and the unit test) at the `instructions` assertion. boot_smoke.mjs: documented its scope (binary /event liveness only; the injection + merge coverage lives in opencode_config.test.ts) so the split is explicit. 80 extension tests green; test:smoke PASS. Closes#25. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… (not skips) The opencode config injection+merge integration test is skipIf(!OC_BIN), and the fast job never fetched the vendored binary before vitest — so it self-skipped in CI. That was the false-green ONE level up: even with the test added, a real injection/merge regression would pass CI because the test didn't run. Fetch opencode before `pnpm -r run test` so it executes and gates. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

jack-champagne

Reviewed — right fix, and cleaner than re-deriving the config in the .mjs. Closes the false-green at both levels: (1) the injection assertion (instructions land + permission merge) now exists in opencode_config.test.ts against the real buildOpencodeConfigContent (no drift), and (2) CI fetches opencode before the suite so the OC_BIN-gated integration test actually runs instead of self-skipping — the deeper false-green. Folds in the #22 permission-merge check and adds global-model preservation (Q129). debug-config vs live GET /config is a fine call (same resolution path). Verified green.