🛟 Updated.
483fde8 parent: cb42032 added
.tickets/issues/0056-blog-post-idea-how-to-query-a-decision-model-fro.yaml +13 -0 | new file mode 100644 | ||
| @@ -0,0 +1,13 @@ | ||
| 1 | +# Managed by IssueSpec. Hand edits are welcome; keep the schema valid. | |
| 2 | +id: 56 | |
| 3 | +title: "Blog post idea: how to query a decision model from a coding agent (skill?)" | |
| 4 | +state: open | |
| 5 | +author: | |
| 6 | + name: k33g | |
| 7 | + email: ph.charriere@gmail.com | |
| 8 | +createdAt: 2026-09-30T06:50:30.091Z | |
| 9 | +updatedAt: 2026-09-30T06:50:30.091Z | |
| 10 | +labels: [] | |
| 11 | +body: "" | |
| 12 | +tasks: [] | |
| 13 | +comments: [] | |
| new file mode 100644 | |||
| @@ -0,0 +1,13 @@ | |||
| 1 | +# Managed by IssueSpec. Hand edits are welcome; keep the schema valid. | ||
| 2 | +id: 56 | ||
| 3 | +title: "Blog post idea: how to query a decision model from a coding agent (skill?)" | ||
| 4 | +state: open | ||
| 5 | +author: | ||
| 6 | + name: k33g | ||
| 7 | + email: ph.charriere@gmail.com | ||
| 8 | +createdAt: 2026-09-30T06:50:30.091Z | ||
| 9 | +updatedAt: 2026-09-30T06:50:30.091Z | ||
| 10 | +labels: [] | ||
| 11 | +body: "" | ||
| 12 | +tasks: [] | ||
| 13 | +comments: [] | ||
added
.tickets/issues/0057-watchdog-kills-the-turn-while-a-tool-runs.yaml +110 -0 | new file mode 100644 | ||
| @@ -0,0 +1,110 @@ | ||
| 1 | +# Managed by IssueSpec. Hand edits are welcome; keep the schema valid. | |
| 2 | +id: 57 | |
| 3 | +title: Watchdog kills the turn while a tool runs or a permission is pending | |
| 4 | +state: open | |
| 5 | +author: | |
| 6 | + name: k33g | |
| 7 | + email: ph.charriere@gmail.com | |
| 8 | +createdAt: 2026-09-30T11:38:05.000Z | |
| 9 | +updatedAt: 2026-09-30T11:38:05.000Z | |
| 10 | +labels: | |
| 11 | + - kind::bug | |
| 12 | + - priority::medium | |
| 13 | +related: | |
| 14 | + - 53 | |
| 15 | +body: | | |
| 16 | + Symptom: "Error: Internal error: generation timed out (stalled)" in an | |
| 17 | + editor (ACP), after a write_file whose permission dialog stayed open; the | |
| 18 | + session records "[write_file rejected by the user]" although the user | |
| 19 | + rejected nothing (session 20260930-072059-d23b9ad5, Gemma 4 26B on DMR). | |
| 20 | + | |
| 21 | + Cause: the watchdog (internal/engine/progress.go) keeps counting while a | |
| 22 | + tool runs. Tools get the watchdog's context (genCtx) and nothing pauses it | |
| 23 | + in WrapTool, so anything longer than watchdogTimeout kills the turn: | |
| 24 | + - a permission dialog answered slowly (ACP and window mode); | |
| 25 | + - a legitimately long tool: bash's budget is 30 s, the default | |
| 26 | + watchdogTimeout is 20 s (30 s in agent.dmr.yaml). | |
| 27 | + | |
| 28 | + Reproduced 2026-09-30 with a fake OpenAI server + scripted ACP client, | |
| 29 | + watchdogTimeout 3s: permission answered after 1 s -> OK; after 6 s -> the | |
| 30 | + exact error, file not written; bash "sleep 6" -> same. Identical on the | |
| 31 | + v0.2.3 and v0.2.4 binaries: not a regression of 0.2.4 (ticket 53 does not | |
| 32 | + touch this path). | |
| 33 | + | |
| 34 | + Workaround: raise watchdogTimeout in the config, or choose "Allow <tool> | |
| 35 | + for this session". | |
| 36 | + | |
| 37 | + Do NOT simply pause the watchdog during tools: today it is the only bound | |
| 38 | + for lsp requests (tool ctx, no timeout of their own) and stdio MCP calls | |
| 39 | + (unbounded unless the server has a timeout in the config). Own bounds: | |
| 40 | + bash 30 s + 2 s WaitDelay, fetch_content 30 s, MCP over HTTP 30 s. | |
| 41 | + The model <-> tools loop is bounded by maxTurns and the loop detector, not | |
| 42 | + by the watchdog, so no endless-loop risk from this change. | |
| 43 | + | |
| 44 | + Proposed design: | |
| 45 | + 1. While a tool runs, replace the watchdog by a tool budget: new config key | |
| 46 | + toolTimeout (default ~2 min), applied in WrapTool on a context that is a | |
| 47 | + child of the TURN's context (Stop still cancels) but not of the watchdog. | |
| 48 | + A tool over budget answers the model with a text error, like bash's | |
| 49 | + "[error: … timeout exceeded]"; the turn goes on. | |
| 50 | + 2. Do not count the permission wait (user's thinking time) in either | |
| 51 | + budget. Optional: a very large cap (e.g. 30 min) for an editor that | |
| 52 | + never answers. | |
| 53 | + 3. When the tool returns, re-arm the watchdog with pl.expect() (the next | |
| 54 | + model call gets firstTokenTimeout). | |
| 55 | +tasks: | |
| 56 | + - id: 1 | |
| 57 | + title: "Failing test first: a tool that sleeps longer than watchdogTimeout must not stall the turn" | |
| 58 | + state: open | |
| 59 | + priority: none | |
| 60 | + author: | |
| 61 | + name: k33g | |
| 62 | + email: ph.charriere@gmail.com | |
| 63 | + createdAt: 2026-09-30T11:38:05.000Z | |
| 64 | + updatedAt: 2026-09-30T11:38:05.000Z | |
| 65 | + - id: 2 | |
| 66 | + title: "toolTimeout config key (default, validation, YAML comments in agent.*.yaml and config.samples)" | |
| 67 | + state: open | |
| 68 | + priority: none | |
| 69 | + author: | |
| 70 | + name: k33g | |
| 71 | + email: ph.charriere@gmail.com | |
| 72 | + createdAt: 2026-09-30T11:38:05.000Z | |
| 73 | + updatedAt: 2026-09-30T11:38:05.000Z | |
| 74 | + - id: 3 | |
| 75 | + title: "WrapTool: tool budget on a turn-child context, watchdog re-armed after the tool" | |
| 76 | + state: open | |
| 77 | + priority: none | |
| 78 | + author: | |
| 79 | + name: k33g | |
| 80 | + email: ph.charriere@gmail.com | |
| 81 | + createdAt: 2026-09-30T11:38:05.000Z | |
| 82 | + updatedAt: 2026-09-30T11:38:05.000Z | |
| 83 | + - id: 4 | |
| 84 | + title: "Exclude the permission wait (ACP and window sinks) from both budgets; decide on a cap" | |
| 85 | + state: open | |
| 86 | + priority: none | |
| 87 | + author: | |
| 88 | + name: k33g | |
| 89 | + email: ph.charriere@gmail.com | |
| 90 | + createdAt: 2026-09-30T11:38:05.000Z | |
| 91 | + updatedAt: 2026-09-30T11:38:05.000Z | |
| 92 | + - id: 5 | |
| 93 | + title: "ACP end-to-end check: permission answered after 2x watchdogTimeout -> file written" | |
| 94 | + state: open | |
| 95 | + priority: none | |
| 96 | + author: | |
| 97 | + name: k33g | |
| 98 | + email: ph.charriere@gmail.com | |
| 99 | + createdAt: 2026-09-30T11:38:05.000Z | |
| 100 | + updatedAt: 2026-09-30T11:38:05.000Z | |
| 101 | + - id: 6 | |
| 102 | + title: "Docs EN+FR: configuration reference (watchdogTimeout vs toolTimeout), explanation" | |
| 103 | + state: open | |
| 104 | + priority: none | |
| 105 | + author: | |
| 106 | + name: k33g | |
| 107 | + email: ph.charriere@gmail.com | |
| 108 | + createdAt: 2026-09-30T11:38:05.000Z | |
| 109 | + updatedAt: 2026-09-30T11:38:05.000Z | |
| 110 | +comments: [] | |
| new file mode 100644 | |||
| @@ -0,0 +1,110 @@ | |||
| 1 | +# Managed by IssueSpec. Hand edits are welcome; keep the schema valid. | ||
| 2 | +id: 57 | ||
| 3 | +title: Watchdog kills the turn while a tool runs or a permission is pending | ||
| 4 | +state: open | ||
| 5 | +author: | ||
| 6 | + name: k33g | ||
| 7 | + email: ph.charriere@gmail.com | ||
| 8 | +createdAt: 2026-09-30T11:38:05.000Z | ||
| 9 | +updatedAt: 2026-09-30T11:38:05.000Z | ||
| 10 | +labels: | ||
| 11 | + - kind::bug | ||
| 12 | + - priority::medium | ||
| 13 | +related: | ||
| 14 | + - 53 | ||
| 15 | +body: | | ||
| 16 | + Symptom: "Error: Internal error: generation timed out (stalled)" in an | ||
| 17 | + editor (ACP), after a write_file whose permission dialog stayed open; the | ||
| 18 | + session records "[write_file rejected by the user]" although the user | ||
| 19 | + rejected nothing (session 20260930-072059-d23b9ad5, Gemma 4 26B on DMR). | ||
| 20 | + | ||
| 21 | + Cause: the watchdog (internal/engine/progress.go) keeps counting while a | ||
| 22 | + tool runs. Tools get the watchdog's context (genCtx) and nothing pauses it | ||
| 23 | + in WrapTool, so anything longer than watchdogTimeout kills the turn: | ||
| 24 | + - a permission dialog answered slowly (ACP and window mode); | ||
| 25 | + - a legitimately long tool: bash's budget is 30 s, the default | ||
| 26 | + watchdogTimeout is 20 s (30 s in agent.dmr.yaml). | ||
| 27 | + | ||
| 28 | + Reproduced 2026-09-30 with a fake OpenAI server + scripted ACP client, | ||
| 29 | + watchdogTimeout 3s: permission answered after 1 s -> OK; after 6 s -> the | ||
| 30 | + exact error, file not written; bash "sleep 6" -> same. Identical on the | ||
| 31 | + v0.2.3 and v0.2.4 binaries: not a regression of 0.2.4 (ticket 53 does not | ||
| 32 | + touch this path). | ||
| 33 | + | ||
| 34 | + Workaround: raise watchdogTimeout in the config, or choose "Allow <tool> | ||
| 35 | + for this session". | ||
| 36 | + | ||
| 37 | + Do NOT simply pause the watchdog during tools: today it is the only bound | ||
| 38 | + for lsp requests (tool ctx, no timeout of their own) and stdio MCP calls | ||
| 39 | + (unbounded unless the server has a timeout in the config). Own bounds: | ||
| 40 | + bash 30 s + 2 s WaitDelay, fetch_content 30 s, MCP over HTTP 30 s. | ||
| 41 | + The model <-> tools loop is bounded by maxTurns and the loop detector, not | ||
| 42 | + by the watchdog, so no endless-loop risk from this change. | ||
| 43 | + | ||
| 44 | + Proposed design: | ||
| 45 | + 1. While a tool runs, replace the watchdog by a tool budget: new config key | ||
| 46 | + toolTimeout (default ~2 min), applied in WrapTool on a context that is a | ||
| 47 | + child of the TURN's context (Stop still cancels) but not of the watchdog. | ||
| 48 | + A tool over budget answers the model with a text error, like bash's | ||
| 49 | + "[error: … timeout exceeded]"; the turn goes on. | ||
| 50 | + 2. Do not count the permission wait (user's thinking time) in either | ||
| 51 | + budget. Optional: a very large cap (e.g. 30 min) for an editor that | ||
| 52 | + never answers. | ||
| 53 | + 3. When the tool returns, re-arm the watchdog with pl.expect() (the next | ||
| 54 | + model call gets firstTokenTimeout). | ||
| 55 | +tasks: | ||
| 56 | + - id: 1 | ||
| 57 | + title: "Failing test first: a tool that sleeps longer than watchdogTimeout must not stall the turn" | ||
| 58 | + state: open | ||
| 59 | + priority: none | ||
| 60 | + author: | ||
| 61 | + name: k33g | ||
| 62 | + email: ph.charriere@gmail.com | ||
| 63 | + createdAt: 2026-09-30T11:38:05.000Z | ||
| 64 | + updatedAt: 2026-09-30T11:38:05.000Z | ||
| 65 | + - id: 2 | ||
| 66 | + title: "toolTimeout config key (default, validation, YAML comments in agent.*.yaml and config.samples)" | ||
| 67 | + state: open | ||
| 68 | + priority: none | ||
| 69 | + author: | ||
| 70 | + name: k33g | ||
| 71 | + email: ph.charriere@gmail.com | ||
| 72 | + createdAt: 2026-09-30T11:38:05.000Z | ||
| 73 | + updatedAt: 2026-09-30T11:38:05.000Z | ||
| 74 | + - id: 3 | ||
| 75 | + title: "WrapTool: tool budget on a turn-child context, watchdog re-armed after the tool" | ||
| 76 | + state: open | ||
| 77 | + priority: none | ||
| 78 | + author: | ||
| 79 | + name: k33g | ||
| 80 | + email: ph.charriere@gmail.com | ||
| 81 | + createdAt: 2026-09-30T11:38:05.000Z | ||
| 82 | + updatedAt: 2026-09-30T11:38:05.000Z | ||
| 83 | + - id: 4 | ||
| 84 | + title: "Exclude the permission wait (ACP and window sinks) from both budgets; decide on a cap" | ||
| 85 | + state: open | ||
| 86 | + priority: none | ||
| 87 | + author: | ||
| 88 | + name: k33g | ||
| 89 | + email: ph.charriere@gmail.com | ||
| 90 | + createdAt: 2026-09-30T11:38:05.000Z | ||
| 91 | + updatedAt: 2026-09-30T11:38:05.000Z | ||
| 92 | + - id: 5 | ||
| 93 | + title: "ACP end-to-end check: permission answered after 2x watchdogTimeout -> file written" | ||
| 94 | + state: open | ||
| 95 | + priority: none | ||
| 96 | + author: | ||
| 97 | + name: k33g | ||
| 98 | + email: ph.charriere@gmail.com | ||
| 99 | + createdAt: 2026-09-30T11:38:05.000Z | ||
| 100 | + updatedAt: 2026-09-30T11:38:05.000Z | ||
| 101 | + - id: 6 | ||
| 102 | + title: "Docs EN+FR: configuration reference (watchdogTimeout vs toolTimeout), explanation" | ||
| 103 | + state: open | ||
| 104 | + priority: none | ||
| 105 | + author: | ||
| 106 | + name: k33g | ||
| 107 | + email: ph.charriere@gmail.com | ||
| 108 | + createdAt: 2026-09-30T11:38:05.000Z | ||
| 109 | + updatedAt: 2026-09-30T11:38:05.000Z | ||
| 110 | +comments: [] | ||