Skip to content

fix/swarm-model-sentinel-allowlist - #1

Draft
gdamprint-cmyk wants to merge 1708 commits into
masterfrom
fix/swarm-model-sentinel-allowlist
Draft

gdamprint-cmyk wants to merge 1708 commits into
masterfrom
fix/swarm-model-sentinel-allowlist

Conversation

@gdamprint-cmyk

Copy link
Copy Markdown
Owner

This draft pull request was opened by Entire to back a trail. Feel free to edit the title or body, but keep the hidden marker below: Entire uses it to close this pull request when the trail closes.

jcode and others added 30 commits November 6, 2026 15:52
Entire-Checkpoint: 01M6RCW26HTG3D1AG06AKRFV08
Entire-Checkpoint: 01M6RG8ETEZYR68JGVS6Z9BKJB
Entire-Checkpoint: 01M6RJ08BGSA3P8CFYCA5JGQZA
…build

Entire-Checkpoint: 01M6RNAJKMY06EECE6VY8FFVKY
…the agent

Entire-Checkpoint: 01M6RWSQW4MWDA3VGEMC3B02XT
Entire-Checkpoint: 01M6SZY5KHE8WH1JCF3EHKXRB1
Entire-Checkpoint: 01M6T36Z07H8Y5B8R7B9HX3KFC
…able-model-invocation

Port the Hermes L1 skill-creation loop and the L4 catalog-hygiene gap.

Entire-Checkpoint: 01M6T3BVJRFYGKXAD3Y6PXY8CS
Entire-Checkpoint: 01M6T4F28MQQJVF2Q5GD6CH88D
…pact the todo board, and repair agentgrep scoping

Entire-Checkpoint: 01M6XGXJJCDWWCQZBAQV1KAMCC
…l state

Entire-Checkpoint: 01M6XHS48A187TGV5YJVKGAYDQ
Entire-Checkpoint: 01M6XKDWSZJV8A2RYH3HPTF148
Entire-Checkpoint: 01M6XKFHQRT4DN4MEPBR0PHQXC
…ude hook

Entire-Checkpoint: 01M6XKTDQCG1NNVFXYJXYYW038
Entire-Checkpoint: 01M6XNE5KHJ3Q1K4HZMYGXXMQ8
… board

The todo tool's model-facing output carried every tool-maintained history
array: `confidence_history` on each card and `understands_user_intent_history`
on the plan. These are append-only trails the tool owns, one entry per
observation, so they grow without bound as a board is worked on. The model
never writes them and never reads them for a decision.

The gates that use them (`is_spike`, the completion report) read the STORED
`TodoItem`/`TodoPlan` structs directly, so nothing about them changes. The
stored state and the tool metadata keep every entry; only the model-facing
text is compacted.

The current `confidence` and `completion_confidence` states stay: they are
single values, cheap, and the model needs them to judge the work.
`user_intention` and `understands_user_intent` stay for the same reason.

Measured on a 100-card board with a 5-entry confidence history per card and
a 5-entry plan intent history: 77695 bytes before, 64578 bytes after
(13117 bytes, 16.9% saved). The saving grows with the number of observations.

Tests: `cargo test -p jcode-app-core --lib todo::` -> 99 passed, 0 failed.
`cargo check -p jcode-app-core --tests` clean.

Entire-Checkpoint: 01M6XQ6KGAQWRE3CR2519863MB
The doc comment on `public_todos_compact` listed the two card- and plan-level
history arrays that the model-facing rendering drops, but not the five
goal-level ones (`closed_feedback_loop_history`,
`feedback_loop_relevance_history`, `feedback_loop_coverage_history`,
`feedback_loop_traceability_history`, `delivery_state_history`). They are
stripped by the same recursive pass, so the comment now says so.

This is a comment-only change: no behaviour differs from 10911e5.

Entire-Checkpoint: 01M6XQH2R4T7VK9PD3Y8FJ2C6X
Entire-Checkpoint: 01M6XSY3957WTVEAC3AAQ5HHWM
Entire-Checkpoint: 01M6Y6XJ67D5GVQBCP6VP2VFWA
Entire-Checkpoint: 01M6YHT2VB8ME01XVP7PVZNNMX
Entire-Checkpoint: 01M6YPFTY05ESX7BYJV014X1P9
…obe path

Entire-Checkpoint: 01M6YZE8G0PBAJWAZH2ZMEBN8R
Entire-Checkpoint: 01M6Z1R69Q3M4ZQDQN7PAA2W50
Entire-Checkpoint: 01M6Z8DGRX1TE4CECQWMGM3TJV
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants