Skill: inform the agent of its own classifier

coilyco-ops-gaming commented

2026-08-17 17:12:58 +00:00

Member

The agent cannot currently describe the classifier or routing layer that governs its inputs and actions. It sees only the fixed policy layer (trust split, guardfile tool grants, house style) and has no inspectable source for the decisions behind what it receives, what it may act on, and whether it replies, reacts, or stays silent.

Gap: no skill or harness-provided explanation informs the agent of its classifier's decision basis, including admitted speaker class, directness, scope and authorization checks, and the reply/react/silent decision with reason codes.

Desired outcome: a skill the agent can load that describes its own classifier accurately. Policy logic stays in the harness; the agent receives an inspectable, non-bypassable explanation rather than guessing from the context it was handed.

The agent cannot currently describe the classifier or routing layer that governs its inputs and actions. It sees only the fixed policy layer (trust split, guardfile tool grants, house style) and has no inspectable source for the decisions behind what it receives, what it may act on, and whether it replies, reacts, or stays silent. Gap: no skill or harness-provided explanation informs the agent of its classifier's decision basis, including admitted speaker class, directness, scope and authorization checks, and the reply/react/silent decision with reason codes. Desired outcome: a skill the agent can load that describes its own classifier accurately. Policy logic stays in the harness; the agent receives an inspectable, non-bypassable explanation rather than guessing from the context it was handed.

coilyco-ops-gaming added the

🔒⚠️📦⚠️🔒 SANDBOXED 🔒⚠️📦⚠️🔒

label

2026-08-17 17:12:58 +00:00

coilyco-ops added the

labels

2026-08-17 17:18:10 +00:00

coilyco-ops referenced this issue

2026-08-17 17:18:33 +00:00

Deep: harness classifier awareness note #907

coilyco-ops referenced this issue

2026-08-17 18:26:01 +00:00

Skills are skill-shaped directories, not skills: every reference file is inlined into every prompt #859

coilyco-ops referenced this issue

2026-08-17 19:51:58 +00:00

Skills are skill-shaped directories, not skills: every reference file is inlined into every prompt #859

coilyco-ops referenced this issue

2026-08-17 20:32:00 +00:00

Confirm wildcard access for Deep to respond to everyone in the Discord #909

coilysiren removed the

🔒⚠️📦⚠️🔒 SANDBOXED 🔒⚠️📦⚠️🔒

label

2026-08-18 22:10:19 +00:00

coilyco-ops referenced this issue from a commit

2026-08-19 04:46:20 +00:00

feat(skills): describe the harness gates the model cannot see

coilyco-ops commented

2026-08-19 04:48:31 +00:00

Member

Landed in 61a8d0d as coilyco-general/references/harness-gates.md.

Written from the code rather than from this issue's description of it. The content-gate half comes from agent/content-classes.yaml's own header, the summoning half from agent.go.

What it gives the agent an inspectable source for:

The gate is allow-by-default, with denied classes as the exception, and the allowed classes enumerated so an ordinary request has somewhere to go rather than being forced into a refusal bucket.
A sensitive class changes the shape of a refusal, not the verdict — a generic redirect naming no category, because naming the rule that fired tells a member what to avoid saying next time.
Sensitive wins ties.
Every block is visible, since silent no-reply was rejected, so a refusal is never mistakable for an outage.
Summoning: mention, reply to its own message, or a message in a thread it opened. Another bot is refused unless allowlisted, and its own messages are dropped first.

The useful part is the shape rather than the class list. A refusal that names no reason is the sensitive branch working as designed, not evasion and not a fault, and a seat that knows this can explain a refusal it did not itself make instead of guessing at it or apologising for it.

Where I did not follow the issue

This asks for reason codes among the decision basis. I did not find any and did not invent them, so they are absent from the file rather than described. If they exist somewhere I did not read, that is a gap in the file worth filling.

Fetchable, not inline

Deliberate, and it costs zero prompt bytes. This is explanatory knowledge that matters when the subject comes up, which is exactly the progressive-disclosure case measured on #968: the Dowel root is 6 of 7 inline at 24KB a turn because every author believed their file was the one that must not be missed. Adding a ninth inline file to fix a self-description gap would have been that same move.

Reach, and it is a real limitation

coilyco-general is loaded by Deep, Dowel, and the HTTP profile. Echo does not load it, and Echo is the lane whose content gate fires most. Fixing that needs either a shared root or duplication, neither of which is freeze work. Flagged rather than worked around.

#906 stays open on the runtime half, which this does not supply and says so.

Refs #906, #968, #227, #1011

**Landed in `61a8d0d`** as `coilyco-general/references/harness-gates.md`. Written from the code rather than from this issue's description of it. The content-gate half comes from `agent/content-classes.yaml`'s own header, the summoning half from `agent.go`. What it gives the agent an inspectable source for: * **The gate is allow-by-default**, with denied classes as the exception, and the allowed classes enumerated so an ordinary request has somewhere to go rather than being forced into a refusal bucket. * **A `sensitive` class changes the shape of a refusal, not the verdict** — a generic redirect naming no category, because naming the rule that fired tells a member what to avoid saying next time. * **Sensitive wins ties.** * **Every block is visible**, since silent no-reply was rejected, so a refusal is never mistakable for an outage. * **Summoning**: mention, reply to its own message, or a message in a thread it opened. Another bot is refused unless allowlisted, and its own messages are dropped first. The useful part is the shape rather than the class list. A refusal that names no reason is **the sensitive branch working as designed**, not evasion and not a fault, and a seat that knows this can explain a refusal it did not itself make instead of guessing at it or apologising for it. ## Where I did not follow the issue This asks for reason codes among the decision basis. **I did not find any and did not invent them**, so they are absent from the file rather than described. If they exist somewhere I did not read, that is a gap in the file worth filling. ## Fetchable, not inline Deliberate, and it costs zero prompt bytes. This is explanatory knowledge that matters when the subject comes up, which is exactly the progressive-disclosure case measured on #968: the Dowel root is 6 of 7 inline at 24KB a turn because every author believed their file was the one that must not be missed. Adding a ninth inline file to fix a self-description gap would have been that same move. ## Reach, and it is a real limitation `coilyco-general` is loaded by Deep, Dowel, and the HTTP profile. **Echo does not load it**, and Echo is the lane whose content gate fires most. Fixing that needs either a shared root or duplication, neither of which is freeze work. Flagged rather than worked around. #906 stays open on the runtime half, which this does not supply and says so. Refs #906, #968, #227, #1011

coilyco-ops referenced this issue

2026-08-19 04:49:06 +00:00

sirens echo should have harness level enforcement of the type of content it will reply to, call it content classifier #227

coilyco-ops referenced this issue

2026-08-19 04:49:22 +00:00

Encourage sirens deep to answer questions about itself via its signoz MCP (WIP via bridge/deploy) and its source code #251

Rows
Columns

Skill: inform the agent of its own classifier #908

Where I did not follow the issue

Fetchable, not inline

Reach, and it is a real limitation