Skip to content

context-forge 0.10.0: How to Act — list assumptions up front, harden against rationalization - #23

Merged
davcs86 merged 2 commits into
mainfrom
ccr-f6590c54-6qjzub-context-forge
Oct 18, 2026
Merged

davcs86 merged 2 commits into
mainfrom
ccr-f6590c54-6qjzub-context-forge

Conversation

@davcs86

@davcs86 davcs86 commented Oct 18, 2026 •

Copy link
Copy Markdown
Owner

Summary

Companion to #22 (design-buddy DF-7). This PR changes only the generated ## How to Act behavioral contract that /context-constitution prepends to a host repo's root CLAUDE.md. context-forge's own skill gates and governance (CF-*) are not changed.

  1. Assumptions are listed up front (behavior feat: scaffold Claude Artifacts portfolio app #1). "Before acting, always list every assumption you are making on your own — scope, interpretation, defaults, a choice between options — each with its basis, or say "none"."
  2. Rationalization hardening:
    • Inline red flags: each of the four behaviors gets a Red flag: clause quoting the phrase that usually comes right before that behavior is broken ("too small to ask", "might need it later", "cleaner if I also…", "should work").
    • Evidence before "done" (fix(design-buddy): repair unparseable SKILL frontmatter + align with marketplace docs #4): the agent may claim done only with the check's output from this session in hand, instead of an unverifiable "loop until verified".
    • No-exceptions closer: a final line removes the small/obvious-task exemption and the letter-vs-intent loophole.

Design notes

  • The four-behavior shape is kept. CF-N2, config-protocol.md and the README all assume exactly four behaviors, so the red flags go inline instead of in a separate table.
  • citeIds tails need no change: the "(enforced by …)" tails still append to each behavior.
  • Re-runs are safe. The sentinel markers are unchanged, so re-running /context-constitution replaces the block in place (CF-4).
  • Cost: about five more lines in an auto-loaded block. That is an accepted CF-N4 trade for behavior that holds up under pressure.
  • Version 0.9.0 → 0.10.0 in all three manifests. Pre-1.0, a behavior change takes a minor bump.

Verification

  • python3 scripts/validate_manifests.py --self-test and python3 scripts/validate_manifests.py pass.
  • Every plugins/*/scripts/validate.py --self-test and validator run passes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NWknUUR4cHtnnYiUq41wiy

…ore acting

Extend behavior #1 ("Don't assume — ask") of the generated behavioral
contract: before acting, the agent always lists every assumption it is
making on its own (scope, interpretation, defaults, choices between
options) with its basis, or says "none", so the user can correct one
before work builds on it. Keeps the four-behavior shape (CF-N2) and the
sentinel markers, so re-runs replace the block in place.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NWknUUR4cHtnnYiUq41wiy
Harden the behavioral contract against rationalization: each of the four
behaviors gains an inline *Red flag:* quoting the phrase that precedes its
violation, #4 now requires the check's output from this session before
claiming done, and a closing line removes the small/obvious-task exemption
and letter-vs-intent loophole. Four-behavior shape (CF-N2) and sentinels
unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NWknUUR4cHtnnYiUq41wiy
@davcs86 davcs86 changed the title context-forge 0.10.0: How to Act #1 — list unilateral assumptions before acting context-forge 0.10.0: How to Act — list assumptions up front, harden against rationalization Oct 18, 2026
@davcs86
davcs86 merged commit d9f79bf into main Oct 18, 2026
2 checks passed
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.

2 participants