New: 98,000+ skills indexed from GitHub repos with 100+ stars · every skill safety-checkedBrowse top skillsTop skills
Esc to closeBrowse all skills
Submit a skillBrowse skills
Submit a skillBrowse skills

Agent skill

Call Correction Propagation Auditor

by CALLE-AI in CALLE-AI/awesome-phone-call-agents

Offline experimental CALL-E helper that tracks agent self-corrections mid-call ("Sorry, I said Tuesday; I meant Thursday"), follows each correction chain, and verifies the corrected value - not the…

MITAuto-check passed

Install Call Correction Propagation Auditor

skills CLI
$ npx skills add CALLE-AI/awesome-phone-call-agents --skill call-correction-propagation-auditor -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install CALLE-AI/awesome-phone-call-agents call-correction-propagation-auditor --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/CALLE-AI/awesome-phone-call-agents.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/call-correction-propagation-auditor .claude/skills/call-correction-propagation-auditor && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
call-correction-propagation-auditor
GitHub stars
107
Token cost
~2.3k tokens
SKILL.md length
1,142 words
Files
10 (incl. scripts, references)
Skills in repo
101
Repo updated
First seen
Licence
MIT

At a glance

Offline experimental CALL-E helper that tracks agent self-corrections mid-call ("Sorry, I said Tuesday; I meant Thursday"), follows each correction chain, and verifies the corrected value - not the…

  • Works in 5 steps: Mask first. Any 7+-digit run (separators… → Detect. Agent-side sentences (questions… → Build chains. Events whose old_value… → …
  • SKILL.md covers When To Use, When Not To Use, Verdicts and How It Works, plus 5 more sections
  • Runs Python scripts from its folder; calls python3 and python

What it does

Call Correction Propagation Auditor is an agent skill from CALLE-AI/awesome-phone-call-agents. Offline experimental CALL-E helper that tracks agent self-corrections mid-call ("Sorry, I said Tuesday; I meant Thursday"), follows each correction chain, and verifies the corrected value - not the superseded one - reached the postsummary, flagging stale values before automation acts; not semantic repair analysis, proof of deception, or authorization to act.

Its SKILL.md is about 2.3k tokens, which your agent loads only when the skill is triggered. The skill folder holds 11 other files, including scripts and reference files (for example `references/example-call-result-none.json`, `references/example-call-result-stale.json` and `references/example-call-result-unconfirmed.json`).

The repository describes itself as: Portable phone-call Agent Skills, apps, examples, adapters, and scheduler recipes for AI agents. The licence is MIT.

Example prompts

  • “Sorry, I said Tuesday; I meant Thursday”
  • “/call-correction-propagation-auditor”

Requirements

  • Python 3

Workflow steps

5 steps, taken from the first numbered list in SKILL.md.

  1. Mask first. Any 7+-digit run (separators included) is masked in
  2. Detect. Agent-side sentences (questions excluded) are scanned for
  3. Build chains. Events whose old_value supersedes a chain's final
  4. Check propagation. Each chain's final and superseded values are
  5. Confirmation window. A chain is confirmed when any later agent

What it can do on your machine

Read from SKILL.md and the folder at commit 38d4118. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Ships 2 files in scripts/ (Python), which the agent can run.

    Shell commands in SKILL.md call:

    • python3
    • python

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    Links to these hosts (documentation or services it may open):

    • jstor.org
    • aclanthology.org

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Call Correction Propagation Auditor loads about 2.3k tokens when it runs, and up to ~5.3k if it reads all its reference files. Until then it costs about 99 tokens; SKILL.md has 1,142 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~99
When it runs · the whole SKILL.md, loaded when a task matches
~2.3k
With references · SKILL.md plus every file in references/, read only if the agent opens them
~5.3k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); the scripts in this folder are not scanned.

SKILL.md

The full file from CALLE-AI/awesome-phone-call-agents at commit 38d4118, republished under its MIT licence (© CALLE-AI). 1,142 words, ~2,332 tokens.

Download SKILL.mdSave it as .claude/skills/call-correction-propagation-auditor/SKILL.md (or your agent's skills folder). This skill also uses 9 other files; get the full folder from GitHub.
name
call-correction-propagation-auditor
description
Offline experimental CALL-E helper that tracks agent self-corrections mid-call ("Sorry, I said Tuesday; I meant Thursday"), follows each correction chain, and verifies the corrected value - not the superseded one - reached the post_summary, flagging stale values before automation acts; not semantic repair analysis, proof of deception, or authorization to act.
license
MIT
research
arXiv:2310.01798, arXiv:2311.08516, arXiv:2605.06527

call-correction-propagation-auditor

A mid-call correction that never reaches the record is a silent lie the summary tells.

Every CALL-E application downstream of a call parses the agent-written post_summary blindly: dates feed writebacks, times feed calendars, amounts feed charges. Agents correct themselves mid-call all the time - "Sorry, I said Tuesday; I meant Thursday" - and conversation-analysis work shows speakers prefer exactly this self-repair form. But intrinsic self-correction is unreliable: an agent that fixed its own error in speech may still write the pre-correction value into the record, because a later correction does not automatically invalidate the earlier stored value in an LLM's context. The sibling call-post-summary-faithfulness-auditor anchors each summary claim to ANY occurrence in the transcript - so a summary carrying the PRE-correction "Tuesday" still grades SUPPORTED, because "Tuesday" was genuinely spoken. That is this skill's companion blind spot: faithfulness to the transcript is not faithfulness to the correction. This skill follows each agent self-correction chain and verifies that the final corrected value - not the superseded one - reached the summary, before automation acts on the stale one.

When To Use

  • before any writeback, calendar entry, or outcome parse acts on the post_summary of a call where the agent restated or corrected details
  • after any call whose transcript shows the agent revising a date, time, amount, count, or address mid-conversation
  • combined with call-post-summary-faithfulness-auditor: run both - the faithfulness auditor anchors claims correction-agnostically, this skill grades what happened to each correction

When Not To Use

  • when the CALLEE revised their answer ("Actually, make it Friday") - that is answer instability, the object of the provenance-grade skill
  • when the CALLEE initiated the repair ("No, Thursday") - callee-side other-repair signals belong to call-repair-sequence-auditor
  • for semantic repair analysis; marker detection is lexical, so a paraphrased correction can go undetected
  • on masked spans; phone digits are masked before analysis by design and are unverifiable by the same token

Verdicts

VerdictMeaningSuggested routing
STALE_VALUE_IN_SUMMARYa superseded value appears in the summary while the corrected one does notblock the writeback; a human must read the call
CORRECTIONS_UNCONFIRMEDa correction chain has no agent restate or callee ack within the confirmation windowverify the corrected value before acting
PROPAGATEDdetected chains are confirmed and no stale-only summary value was found; summary checks may still be unreported or ambiguousinspect every summary check and advisory; human review before acting
NO_SELF_CORRECTIONSno agent self-correction detected; reason is transcript_missing (no turns) or absent (none detected)no correction audit applies; run the faithfulness auditor

How It Works

Five deterministic stages, offline, no LLM:

  1. Mask first. Any 7+-digit run (separators included) is masked in the summary and every turn before any correction work, keeping the last two characters. Phone numbers can never anchor and never leak into cards.
  2. Detect. Agent-side sentences (questions excluded) are scanned for correction markers - sorry_i_said, meant, not_x_but_y, y_not_x, correction, let_me_correct, incorrect, should_be, my_mistake, actually_its, scratch - the first matching marker wins the sentence. Each marker binds an old and a new value of one kind (date, weekday, clock, money, count, street).
  3. Build chains. Events whose old_value supersedes a chain's final or superseded value extend that chain; otherwise they start a new one
    • "Tuesday -> Thursday -> Friday" becomes one chain with two superseded values.
  4. Check propagation. Each chain's final and superseded values are folded (case, meridiem to 24h, month names, digit commas) and searched in the folded summary with boundary guards ("14:00" does not match "2:14:00", "tuesday" does not match "tuesday's"). Superseded-in-summary-without-final is stale; both is ambiguous; final-only is propagated; neither is unreported.
  5. Confirmation window. A chain is confirmed when any later agent turn restates the final value, or a callee turn within turns i+1..i+2 after the correction repeats it or acknowledges ("ok", "that works", "sounds good", ...). Unconfirmed chains cap the verdict at CORRECTIONS_UNCONFIRMED unless a stale value already blocks it.

The card reports per-event corrections, per-chain summary checks, counts, an advisories list (summary_missing, unreported_chain: <value>), and a fixed honesty disclaimer.

Craft: the correction-discipline goal

bash
python3 scripts/correction_propagation_auditor.py craft

Emits a plan_call goal whose CORRECTION DISCIPLINE block instructs the agent - the moment it corrects any detail - to re-state the corrected value in a full sentence, ask the caller to confirm it, and use only the corrected value thereafter, so the summary this skill later audits carries corrections forward instead of stale values.

Show full SKILL.md (437 more words)Show less

Limitations

  • English correction lexicons only; corrections phrased outside the marker list go undetected
  • spelled-out numbers ("four", "twelve") are not correctable values; this is documented as not graded; never guessed, not silently graded
  • masked digit runs (phones, long numbers) are never correctable by design
  • proximity resolution picks the nearest preceding same-kind value, so disjunctive lists ("Tuesday or Wednesday") may bind the wrong superseded value
  • corrections packed with multiple value pairs in one sentence ("it's the 21st, not the 16th, and 2 bags, not 3") may undercount - each marker binds one old/new pair
  • weekday possessives and plurals ("Tuesday's") are treated as different tokens by the boundary rule and will not match a plain "Tuesday"
  • agent corrections prompted by the CALLEE ("No, Thursday") are other-corrections and out of scope
  • the whole audit is lexical, not semantic; a paraphrased correction or summary restatement can evade detection
  • the aggregate PROPAGATED label does not prove every correction reached the summary; inspect each chain's summary check and the advisory list

Testing

bash
python -m pytest scripts/test_correction_propagation_auditor.py -q
python3 scripts/test_correction_propagation_auditor.py

The suite in scripts/ covers masking, marker detection, chain building, propagation outcomes, the confirmation window, verdict routing, the CLI surface, and craft output against the bundled fixtures.

Scientific Foundation

FunctionCitation
Method anchor (CA self- vs other-repair organization)Schegloff, E. A., Jefferson, G., & Sacks, H. (1977). "The Preference for Self-Correction in the Organization of Repair in Conversation." Language 53(2):361-382. DOI 10.2307/413107, https://www.jstor.org/stable/413107
Failure-mode evidence (intrinsic self-correction unreliable - corrections must be tracked, not assumed to propagate)Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., & Zhou, D. (2024). "Large Language Models Cannot Self-Correct Reasoning Yet." ICLR 2024. arXiv:2310.01798
Craft-mode grounding (correction succeeds when the error location is explicit - hence re-state + readback instruction)Tyen, G., Mansoor, H., Carbune, V., Chen, P., & Mak, T. (2024). "LLMs cannot find reasoning errors, but can correct them given the error location." Findings of ACL 2024. arXiv:2311.08516, https://aclanthology.org/2024.findings-acl.826
Domain + recency (implicit-conflict failure: later observation fails to invalidate earlier stored value - exactly the stale summary value)Chao, H., Bai, Y., Sheng, R., Li, T., & Sun, Y. (2026). "STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?" arXiv:2605.06527

Boundaries

  • the provenance-grade skill grades CALLEE answer instability (its signal I); this skill audits AGENT self-corrections and the summary record - the two directions do not overlap
  • call-repair-sequence-auditor detects CALLEE-initiated other-repair trouble signals; agent self-corrections are this skill's object
  • call-post-summary-faithfulness-auditor anchors summary claims correction-agnostically - any transcript occurrence satisfies it, so a stale summary can still grade FAITHFUL; they compose, run both

This skill is the deterministic, offline, lexical-marker variant of that methodology - not a learned model. Every output is labeled with a fixed disclaimer stating exactly that.

© CALLE-AI, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 9 other files (scripts, references) in skills/call-correction-propagation-auditor of CALLE-AI/awesome-phone-call-agents.

  • SKILL.md
  • references/example-call-result-none.json
  • references/example-call-result-stale.json
  • references/example-call-result-unconfirmed.json
  • references/example-call-result.json
  • references/example-goal.txt
  • references/examples.md
  • references/safety.md
  • scripts/correction_propagation_auditor.py
  • scripts/test_correction_propagation_auditor.py

Open the folder on GitHubat commit 38d4118

Compare with similar skills

Call Correction Propagation Auditor next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Call Correction Propagation Auditor compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Call Correction Propagation Auditor this skillCALLE-AI/awesome-phone-call-agents107—~2.3kAutomated safety check: PassMIT
Cost Trackingaffaan-m/ECC277k1 repos~1.3kAutomated safety check: PassMIT
Correctcursor/plugins11k3 repos~612Automated safety check: PassNone
TrackingBuilderIO/agent-native7.1k—~8kAutomated safety check: PassNone
CorrectionNxcoreAI/EverRoom3k—~290Automated safety check: PassCustom licence
Time Trackingsickn33/agentic-awesome-skills47k1 repos~3.8kAutomated safety check: PassMIT

Similar skills

  • Cost Tracking

    affaan-m/ECC

    Track and report Claude Code token usage, spending, and budgets from the local ECC cost-tracker metrics log.

    277k GitHub starsUsed in 1 repo~1.3k tokens
    AI & LLM EngineeringAuto-check passed

  • Correct

    cursor/plugins

    Official

    Find the mistakes agents keep repeating in this repo and make each one impossible.

    11k GitHub starsUsed in 3 repos~612 tokens
    Auto-check passed

  • Tracking

    BuilderIO/agent-native

    Server-side analytics tracking with pluggable providers. An agent skill from BuilderIO/agent-native.

    7.1k GitHub stars~8k tokensUpdated yesterday
    Backend & APIsAuto-check passed

  • Correction

    NxcoreAI/EverRoom

    Compute Room overview corrections—citation corrections as per-claim edits and general corrections as a single proposal.

    3k GitHub stars~290 tokensUpdated 2 days ago
    Research & ScienceAuto-check passed

  • Time Tracking

    sickn33/agentic-awesome-skills

    Time entry register: employee, project, client, task, hours, billable flag, rate and amount, invoice and approver.

    47k GitHub starsUsed in 1 repo~3.8k tokens
    Productivity & AutomationAuto-check passed

  • Horizon Track

    ruvnet/ruflo

    Track long-horizon objectives across multiple sessions with milestone checkpoints, progress persistence, and drift detection

    74k GitHub stars~744 tokensUpdated yesterday
    Product & Project ManagementAuto-check: notes

More from CALLE-AI/awesome-phone-call-agents

All 101 skills in this repo

  • Accessible Outing Verifier

    CALLE-AI/awesome-phone-call-agents

    Demonstrates advisory accessibility-planning checks with offline fixtures and a proposed bounded CALL-E workflow; use for exploring unknown or qualified venue claims without making calls.

    107 GitHub stars~1.4k tokensUpdated yesterday
    Auto-check passed

  • Ground Truth Gate

    CALLE-AI/awesome-phone-call-agents

    A skill your agent uses when an agent holds some evidence for a physical-world claim but the evidence is broader, narrower, or older than the exact question asked, and it must first decide whether a…

    107 GitHub stars~3.3k tokensUpdated yesterday
    Auto-check passed

  • Is It Accessible

    CALLE-AI/awesome-phone-call-agents

    Call a venue and ask the accessibility questions that matter to one specific person — step-free entry, hearing loop, guide dogs, quiet hours, changing places — then return a per-need verdict backed…

    107 GitHub stars~4.3k tokensUpdated yesterday
    Auto-check passed

  • Landmark Navigation Assist

    CALLE-AI/awesome-phone-call-agents

    Turns a pre-written, building-level location config into a CALL-E outbound phone-call task that guides a delivery driver through the last few hundred metres to a specific building using landmarks…

    107 GitHub stars~1.6k tokensUpdated yesterday
    Auto-check passed

  • Research Gap Call Verifier

    CALLE-AI/awesome-phone-call-agents

    Turn cited business research into a bounded, approval-gated phone-call plan that asks only unresolved factual questions, then reconcile CALL-E-compatible results without treating voicemail, refusal…

    107 GitHub stars~1.7k tokensUpdated yesterday
    Auto-check passed

  • Structured Outcome Followup Call

    CALLE-AI/awesome-phone-call-agents

    Place a goal-driven CALL-E call that collects specific structured answers, score those answers against a deterministic rubric you supply, and conditionally trigger a follow-up action — all runnable…

    107 GitHub stars~1.4k tokensUpdated yesterday
    Auto-check passed

Questions about Call Correction Propagation Auditor

What does Call Correction Propagation Auditor do?

Offline experimental CALL-E helper that tracks agent self-corrections mid-call ("Sorry, I said Tuesday; I meant Thursday"), follows each correction chain, and verifies the corrected value - not the…. Call Correction Propagation Auditor is an agent skill from CALLE-AI/awesome-phone-call-agents. Offline experimental CALL-E helper that tracks agent self-corrections mid-call ("Sorry, I said Tuesday; I meant Thursday"), follows each correction chain, and verifies the corrected value - not the superseded one - reached the postsummary, flagging stale values before automation acts; not semantic repair analysis, proof of deception, or authorization to act.

How do I install Call Correction Propagation Auditor in Claude Code?

Run `npx skills add CALLE-AI/awesome-phone-call-agents --skill call-correction-propagation-auditor -a claude-code`. Or copy the skill folder (skills/call-correction-propagation-auditor in CALLE-AI/awesome-phone-call-agents) into .claude/skills/call-correction-propagation-auditor in your project. Claude Code loads it when a task matches its description.

How do I install Call Correction Propagation Auditor in Codex?

Run `npx skills add CALLE-AI/awesome-phone-call-agents --skill call-correction-propagation-auditor -a codex`. Or copy the skill folder (skills/call-correction-propagation-auditor in CALLE-AI/awesome-phone-call-agents) into .agents/skills/call-correction-propagation-auditor in your project. Codex loads it when a task matches its description.

Can I use Call Correction Propagation Auditor in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add CALLE-AI/awesome-phone-call-agents --skill call-correction-propagation-auditor -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/call-correction-propagation-auditor, .gemini/skills/call-correction-propagation-auditor, .github/skills/call-correction-propagation-auditor and .opencode/skills/call-correction-propagation-auditor in your project.

What does Call Correction Propagation Auditor need to run?

Going by SKILL.md and its folder, Call Correction Propagation Auditor needs Python for the scripts in its folder and the command-line tools its instructions call (python3 and python). Our summary lists: Python 3.

Does Call Correction Propagation Auditor access the network?

SKILL.md names 2 domains. As links in the text: jstor.org and aclanthology.org. This is read from the text; nothing was executed.

Is Call Correction Propagation Auditor safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. The check reads SKILL.md only: the scripts in the folder are not scanned, so read them before running anything.

What licence does Call Correction Propagation Auditor use?

Call Correction Propagation Auditor is published under the MIT licence (declared in SKILL.md). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Call Correction Propagation Auditor use?

About 2.3k tokens (SKILL.md is roughly 9.3k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 2.9k tokens, read only when the agent opens those files.

What are the alternatives to Call Correction Propagation Auditor?

Skills that share tags, products or a category with Call Correction Propagation Auditor: Cost Tracking (affaan-m/ECC, 277k stars), Correct (cursor/plugins, 11k stars), Tracking (BuilderIO/agent-native, 7.1k stars) and Correction (NxcoreAI/EverRoom, 3k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Call Correction Propagation Auditor?

CALLE-AI (a GitHub organization) maintains it in CALLE-AI/awesome-phone-call-agents, which has 107 GitHub stars. The repository holds 101 skills in this directory. The repository was last updated on October 10, 2026.

Source: CALLE-AI/awesome-phone-call-agents on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.