Skip to content

Make Codex and Claude Code subscriptions go further: a local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.

docs: promote recurring workflow feedback to high priority (#147)

澄潭fffea05cf699
diff --git a/.agents/README.md b/.agents/README.md
index a36a5d6c..d95bde1d 100644
--- a/.agents/README.md
+++ b/.agents/README.md
@@ -34,11 +34,13 @@ reconciliation, not deleting the manifest or force-overwriting local edits.
Commit the manifest, generated files and links together so worktrees stay pinned.
The shared [workflow improvement skill](skills/hiroute-workflow-improvement/SKILL.md)
-owns feedback routing, deduplication, stable analysis labels and verification of
-adopted improvements when later work provides comparable context. Issues preserve
-the resumption context across sessions; feedback does not keep an iteration open. Its CLI and shared
-workflow suggestions go upstream; HiRoute policy suggestions stay here. It has no
-Codex host dependency and is exposed to Claude through its own relative link.
+owns feedback routing, deduplication, independent recurrence counting, automatic
+high-priority promotion from the third confirmed occurrence, stable analysis labels
+and verification of adopted improvements when later work provides comparable
+context. Issues preserve the resumption context across sessions; feedback does not
+keep an iteration open. Its CLI and shared workflow suggestions go upstream;
+HiRoute policy suggestions stay here. It has no Codex host dependency and is exposed
+to Claude through its own relative link.
Generated workflow files and their manifest remain owned by issue-spec.
## Codex-only skills
diff --git a/.agents/skills/hiroute-workflow-improvement/SKILL.md b/.agents/skills/hiroute-workflow-improvement/SKILL.md
index f4ab3101..c74652a2 100644
--- a/.agents/skills/hiroute-workflow-improvement/SKILL.md
+++ b/.agents/skills/hiroute-workflow-improvement/SKILL.md
@@ -18,8 +18,10 @@ a separate broad audit or require a feedback report for every task.
Continue the original task wherever possible. Feedback is not a new planning,
review or delivery gate, and creating an Issue does not authorize its implementation.
-HiRoute's repository policy authorizes filing and labeling evidence-backed feedback
-from its work. Honor a user's narrower scope, such as read-only or no publication.
+HiRoute's repository policy authorizes filing, recurrence comments and labeling
+evidence-backed feedback from its work, including automatic promotion at the
+confirmed recurrence threshold below. Honor a user's narrower scope, such as
+read-only or no publication.
Do not infer this standing authorization when copying this skill to another project.
Do not start agents, recurring jobs, workflow upgrades or automatic rule rewrites.
@@ -52,7 +54,8 @@ Never edit generated `issue-spec-*` files to implement upstream suggestions.
## Deduplicate and write
Use the [report and analysis guide](references/feedback.md) for the body fields,
-label definitions, provider commands and time-window queries.
+independent-occurrence counting, label definitions, provider commands and
+time-window queries.
1. Verify the destination repository and active backend with read-only CLI/auth
inspection; use explicit `--repo` on each command. HiRoute and the upstream above
@@ -65,9 +68,13 @@ label definitions, provider commands and time-window queries.
relevant remaining pages or narrow the query; disclose bounded search coverage.
3. Reuse an existing Issue for the same cause and expected correction. Add only new
evidence or a distinct recurrence through `issue-spec comment create`; do not
- repeat an unchanged report. Respect closed/wontfix decisions. A demonstrated
- regression after a fix may warrant a new Issue linked to the old one; do not
- reopen it automatically. Split unrelated causes even if their symptoms match.
+ create another Issue merely because another task or agent discovered it. Count
+ the original observed report as occurrence 1 and each independently evidenced
+ recurrence once. Ordinary discussion, retries and handoffs do not add occurrences.
+ Link the prior evidence and record the confirmed count or lower bound in the new
+ comment. Respect closed/wontfix decisions. A demonstrated regression after a fix
+ may warrant a new Issue linked to the old one; do not reopen it automatically.
+ Split unrelated causes even if their symptoms match.
4. Prepare a self-contained English body with sanitized reproduction, expected and
observed behavior, impact, owning layer, suggested correction and acceptance.
Include the minimal scenario and prerequisites needed by a future session; a
@@ -87,14 +94,19 @@ label definitions, provider commands and time-window queries.
Provision missing labels without overwriting existing definitions. For a reused
Issue with conflicting category labels, resolve ownership before changing them;
do not silently remove others' classification.
+ When recorded evidence on an open, unresolved Issue reaches at least
+ **3 confirmed occurrences**, automatically promote it to `priority/high` using the
+ [counting and promotion rules](references/feedback.md#count-independent-occurrences-and-promote).
+ This is part of recording the feedback, without another approval or periodic job.
7. Read back the Issue URL, body/comment and labels. If a create returns an uncertain
result, search/read before retrying. If labeling fails after creation, retain the
URL and report the missing labels; retry only that step after resolving the
failure. Never create a replacement Issue just to retry labels, switch providers
after a denial, or claim the labeled feedback is complete when it is partial.
-Report the Issue URL, category, create/reuse result and any incomplete labeling.
-If nothing actionable was found, say so without filing placeholder Issues.
+Report the Issue URL, category, create/reuse result, confirmed occurrence count,
+priority action and any incomplete labeling. If nothing actionable was found,
+say so without filing placeholder Issues.
## Review an interval
diff --git a/.agents/skills/hiroute-workflow-improvement/references/feedback.md b/.agents/skills/hiroute-workflow-improvement/references/feedback.md
index 0325e8e5..d8938820 100644
--- a/.agents/skills/hiroute-workflow-improvement/references/feedback.md
+++ b/.agents/skills/hiroute-workflow-improvement/references/feedback.md
@@ -16,6 +16,11 @@ whole feedback set and the category for grouping; do not add per-session, per-ag
or per-week labels. GitHub timestamps and the dated observations below support
interval analysis. Existing unrelated labels remain intact.
+Priority is separate from category. Use `priority/high` for the recurrence promotion
+below; do not invent another improvement-specific priority label. Reuse its existing
+definition. If missing, provision it only when a qualifying promotion needs it, with
+description `High-priority issue` and suggested color `d73a4a`.
+
## Body fields
Use these fields in ordinary Markdown; they are analysis conventions, not typed
@@ -29,8 +34,9 @@ issue-spec artifacts. Omit unavailable revisions with an explicit reason.
for workflow reports, explain the project-independent or project-specific boundary.
- **Observation:** minimal sanitized command/task sequence, expected versus actual
behavior, and concrete impact. Mark hypotheses explicitly.
-- **Recurrence:** observed count and distinct task contexts, or `unknown`; never infer
- frequency from repeated tool retries or count one report per label as occurrences.
+- **Recurrence:** confirmed count or lower bound, distinct task contexts and links to
+ the original evidence, or `unknown` when independence cannot be established. Use
+ the counting rules below rather than comment totals or a previous summary's number.
- **Suggested improvement:** smallest useful correction and any known workaround.
- **Acceptance and resumption:** observable outcome, minimal sanitized scenario,
necessary prerequisites and stable evidence references for a future session.
@@ -39,8 +45,7 @@ issue-spec artifacts. Omit unavailable revisions with an explicit reason.
For new evidence on an existing Issue, append the observation date, source revision,
new reproduction/impact and what differs. Do not overwrite its original evidence.
-Record a recurrence only for a distinct observed occurrence, not another search or
-retry of the same event.
+A distinct occurrence is useful new evidence even when its symptom is unchanged.
When a correction becomes available, append its public fix/release link and revision.
When HiRoute adopts it, append the adoption PR/commit, installed CLI or generated
@@ -50,6 +55,56 @@ these distinct from occurrence counts. A useful review separates reported findin
upstream fixes available, adopted fixes, verified benefit and remaining friction;
Issue open/closed state alone cannot represent that progression.
+## Count independent occurrences and promote
+
+Count evidence for the same unresolved cause across tasks, without resetting at a
+new session or date. The original report counts as **1** when it records observed
+friction. Each later independent encounter counts once, including one by the same
+person or agent in a different task. A new agent, author, session or comment alone
+does not prove independence: require a distinct task/run context, observation date
+and sanitized scenario/impact with applicable source revisions.
+
+Do not count retries, repeated searches, handoffs, copied reports, acknowledgements,
+design discussion, or fix/adoption validation of the original scenario as additional
+occurrences. Legacy comments can count if their evidence establishes a distinct
+encounter; no new template is required retroactively. Deduplicate evidence referenced
+from multiple comments or Issues. Summaries and reported totals point to evidence;
+they are not additional occurrences. If history is incomplete, report a confirmed
+lower bound; promote only when at least 3 distinct occurrences are supported.
+
+Before posting, reread the Issue and relevant comments so a retry or concurrent
+report of the same event does not create another occurrence comment. A new comment
+records the current context, links the earlier counted evidence and states the
+count before and after this observation. For example:
+
+```markdown
+### Independent recurrence
+- Observed on: YYYY-MM-DD (timezone).
+- Source and context: originating task/PR, applicable revisions, and the distinct task sequence.
+- Observation and impact: sanitized expected/actual behavior and concrete cost.
+- Recurrence: 2 confirmed before this observation; at least 3 including this occurrence.
+- Earlier evidence: original Issue body (1); linked independent recurrence comment (2).
+- Priority: recurrence threshold reached; promote to `priority/high`.
+```
+
+When recording a new occurrence on an **open, unresolved** feedback Issue and the
+confirmed count reaches or exceeds **3**, apply `priority/high` automatically after
+the comment is persisted. The original report plus one independent recurrence
+comment plus the current independent encounter qualifies; an Issue with two ordinary
+discussion comments does not. A newly filed report with 3 independently evidenced
+encounters also qualifies. Verify the recorded evidence, current state and labels
+before promotion; reread after writing to confirm the result.
+
+If already high or higher, preserve that priority. On promotion replace only
+`priority/low` and `priority/medium` when present, preserving category and unrelated
+labels. Do not override an explicit maintainer decision against escalation or a
+closed/wontfix decision; record the new evidence and explain why automatic promotion
+was skipped. A newly filed regression after a fix starts its own count and links the
+old Issue; do not inherit its resolved occurrences. Lower counts do not trigger an
+automatic downgrade, and frequency alone does not assign critical priority or a
+delivery deadline. Reconcile uncertain writes by reading the Issue; retry only a
+missing comment or label operation, without counting the same event again.
+
## Resume after a gap
The Issue is the durable handoff, not the original chat, agent or temporary checkout.
@@ -104,6 +159,21 @@ Issue, attach missing labels with `gh issue edit --add-label`, preserving unrela
labels. Inspect the Issue before retrying an uncertain create; if it exists, repair
only missing labels rather than submitting again.
+For a qualifying promotion, ensure `priority/high` exists, then add it. Include
+`--remove-label 'priority/low'` and/or `--remove-label 'priority/medium'` in the same
+edit command only for lower labels actually present:
+
+```sh
+# Create only when missing; preserve any existing definition:
+gh label create 'priority/high' --repo OWNER/REPO --color d73a4a --description 'High-priority issue'
+gh issue edit NUMBER --repo OWNER/REPO --add-label 'priority/high'
+gh issue view NUMBER --repo OWNER/REPO --json url,state,labels
+```
+
+If the recurrence comment succeeds but promotion fails, keep its URL and confirmed
+count, report the pending priority update, and repair only labeling. Do not repost
+the occurrence or claim promotion succeeded from the intended-action line alone.
+
This provider choice covers ordinary feedback, not issue-spec planning artifacts.
For a non-GitHub target use its supported issue-spec Issue/label operations. If
labeling is unavailable, keep the feedback draft or report an already-created Issue
@@ -121,8 +191,10 @@ gh issue list --repo OWNER/REPO --label 'improvement/feedback' --state all --sea
This selects new Issues, not all occurrences. Also discover older Issues updated in
the interval (`updated:START..END`), then read their discussions through issue-spec
and count only explicitly dated observations in the interval. An update can be a
-label edit or closure and is not evidence of recurrence. Page through results when
-limits are reached, or split the date range and deduplicate by repository + number;
+label edit or closure and is not evidence of recurrence. Use the same independence
+rules above; distinguish occurrences in the interval from the cumulative confirmed
+count used for promotion. Page through results when limits are reached, or split
+the date range and deduplicate by repository + number;
state incomplete coverage if it cannot be exhausted. Use the user's timezone and
explicit inclusive start/end dates in the summary. Include relevant fix/adoption
updates from that window and distinguish verified benefits from untested upgrades