fix(chat): pending shell row stops impersonating an error; chips survive a shell-driven session

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.

[Rchari1]

Two reports from the same screenshot, two unrelated causes. Both were untestable where they lived, and both shipped a user-visible bug, so each rule is extracted to a pure module with tests.

1. "A constant error not going away even though my session is progressing"

The shell row's label chain was command ?? title ?? description. A pending bash part has no input.command yet, so it fell through to the model's free-text description and rendered a full prose sentence in the slot where users read a command — and held it for the command's whole duration (1m55s observed).

The text was a paraphrase of agent guidance:

"A local launch is REFUSED by amico-run while this solver is selected (exit 64), so attempting one only wastes a turn."

Nothing had failed. The solve underneath was on iteration 29 with frames on disk, and the solver was piccolo, so the sentence wasn't even applicable. It just looked exactly like a hard error that wouldn't clear.

The description fallback is worth keeping — it's often a useful "Install deps" — so the fix isn't to drop it but to stop it impersonating a command: unwrap quotes, first line only, clamp to 72 chars with an ellipsis.

2. "What happened to the chips at the top"

The rail's session gate was part.tool.startsWith("amicode_"). A session that did its amicode work through the shell therefore showed no chips at all.

That's what happened here: the same session created a problem workspace, wrote a solvespec.json, and drove a solve to iteration 29 — entirely via bash + amico-run, never calling an amicode_* tool. The rail stayed hidden the whole time despite having real entities to describe.

The gate now also matches a shell part whose command drives amicode (amico-run, .amico/problems, amico plan|spec). It stays session-scoped, which is the entire point of the gate — an unrelated session still shows nothing, and there's a test pinning that.

The refetch key is deliberately not broadened. Shell parts don't mutate the problem view through the tool seam, so counting them would refetch the problem on unrelated shell activity. any (the gate) widens; completed (the refetch key) doesn't.

Testing

15 new tests, including the exact reported strings as regression cases: the quoted guidance sentence must clamp and lose its quote, and a shell-launched solve must open the gate while git status must not. Plus the fiddly edges — pending parts with no input, empty-string commands not beating a usable description, exact clamp boundaries.

ui: 318 pass. tsgo clean in ui and app. Both fixes verified present in the built darwin-arm64 binary (the clamp constant and the shell markers; symbol names are minified away).

Note

These are UI-only. Neither is deploy-gated — worth saying because the chips in particular looked like they might be fixed by shipping something, and they weren't: the gate was rejecting a legitimate session shape.